Skin melanoma segmentation algorithm using dual-channel efficient CNN network

Abstract

Except for early surgical resection, melanoma lacks special treatment, while image segmentation can effectively assist doctors to enhance the efficiency of early diagnosis of melanoma. Due to the non-uniform size, shape and color of melanoma, it is difficult to segment the boundary of its lesion area. To solve the above problems, an improved DC-Unet network segmentation algorithm is proposed in this paper. A channel attention ECA-NET module was first introduced to make the model more focused on the lesion area of melanoma. Finally, the segmentation results are post-processed by Conditional Random Field (CRF) and Test Data Augmentation (TTA) to further refine the segmentation results. The experimental results showed that compared with the DC-Unet algorithm on the ISIC2017, ISIC2018 datasets, the segmentation accuracy was increased from 0.9513, 0.9444 to 0.9623, 0.9537 respectively.